1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 * Ceph - scalable distributed file system
6 * Copyright (C) 2015 Red Hat
8 * This is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License version 2.1, as published by the Free Software
11 * Foundation. See file COPYING.
18 #include "common/DecayCounter.h"
19 #include "common/LogClient.h"
20 #include "common/Timer.h"
21 #include "common/TrackedOp.h"
23 #include "messages/MCommand.h"
26 #include "DamageTable.h"
28 #include "SessionMap.h"
32 #include "PurgeQueue.h"
33 #include "osdc/Journaler.h"
35 // Full .h import instead of forward declaration for PerfCounter, for the
36 // benefit of those including this header and using MDSRank::logger
37 #include "common/perf_counters.h"
53 l_mds_inodes_pin_tail
,
56 l_mds_inodes_with_caps
,
61 l_mds_traverse_forward
,
62 l_mds_traverse_discover
,
63 l_mds_traverse_dir_fetch
,
64 l_mds_traverse_remote_ino
,
67 l_mds_dispatch_queue_len
,
69 l_mds_exported_inodes
,
71 l_mds_imported_inodes
,
97 struct heartbeat_handle_d
;
108 class MDSTableServer
;
109 class MDSTableClient
;
118 * The public part of this class's interface is what's exposed to all
119 * the various subsystems (server, mdcache, etc), such as pointers
120 * to the other subsystems, and message-sending calls.
124 const mds_rank_t whoami
;
126 // Incarnation as seen in MDSMap at the point where a rank is
131 mds_rank_t
get_nodeid() const { return whoami
; }
132 int64_t get_metadata_pool();
134 // Reference to global MDS::mds_lock, so that users of MDSRank don't
135 // carry around references to the outer MDS, and we can substitute
136 // a separate lock here in future potentially.
139 class CephContext
*cct
;
141 bool is_daemon_stopping() const;
143 // Reference to global cluster log client, just to avoid initialising
144 // a separate one here.
147 // Reference to global timer utility, because MDSRank and MDSDaemon
148 // currently both use the same mds_lock, so it makes sense for them
161 MDBalancer
*balancer
;
162 ScrubStack
*scrubstack
;
163 DamageTable damage_table
;
168 SnapServer
*snapserver
;
169 SnapClient
*snapclient
;
171 MDSTableClient
*get_table_client(int t
);
172 MDSTableServer
*get_table_server(int t
);
174 SessionMap sessionmap
;
175 Session
*get_session(client_t client
) {
176 return sessionmap
.get_session(entity_name_t::CLIENT(client
.v
));
179 PerfCounters
*logger
, *mlogger
;
180 OpTracker op_tracker
;
182 // The last different state I held before current
183 MDSMap::DaemonState last_state
;
184 // The state assigned to me by the MDSMap
185 MDSMap::DaemonState state
;
187 bool cluster_degraded
;
189 MDSMap::DaemonState
get_state() const { return state
; }
190 MDSMap::DaemonState
get_want_state() const { return beacon
.get_want_state(); }
192 bool is_creating() const { return state
== MDSMap::STATE_CREATING
; }
193 bool is_starting() const { return state
== MDSMap::STATE_STARTING
; }
194 bool is_standby() const { return state
== MDSMap::STATE_STANDBY
; }
195 bool is_replay() const { return state
== MDSMap::STATE_REPLAY
; }
196 bool is_standby_replay() const { return state
== MDSMap::STATE_STANDBY_REPLAY
; }
197 bool is_resolve() const { return state
== MDSMap::STATE_RESOLVE
; }
198 bool is_reconnect() const { return state
== MDSMap::STATE_RECONNECT
; }
199 bool is_rejoin() const { return state
== MDSMap::STATE_REJOIN
; }
200 bool is_clientreplay() const { return state
== MDSMap::STATE_CLIENTREPLAY
; }
201 bool is_active() const { return state
== MDSMap::STATE_ACTIVE
; }
202 bool is_stopping() const { return state
== MDSMap::STATE_STOPPING
; }
203 bool is_any_replay() const { return (is_replay() || is_standby_replay()); }
204 bool is_stopped() const { return mdsmap
->is_stopped(whoami
); }
205 bool is_cluster_degraded() const { return cluster_degraded
; }
207 void handle_write_error(int err
);
209 void handle_conf_change(const struct md_config_t
*conf
,
210 const std::set
<std::string
> &changed
)
212 purge_queue
.handle_conf_change(conf
, changed
, *mdsmap
);
215 void update_mlogger();
217 // Flag to indicate we entered shutdown: anyone seeing this to be true
218 // after taking mds_lock must drop out.
221 // PurgeQueue is only used by StrayManager, but it is owned by MDSRank
222 // because its init/shutdown happens at the top level.
223 PurgeQueue purge_queue
;
225 class ProgressThread
: public Thread
{
229 explicit ProgressThread(MDSRank
*mds_
) : mds(mds_
) {}
230 void * entry() override
;
232 void signal() {cond
.Signal();}
235 list
<Message
*> waiting_for_nolaggy
;
236 list
<MDSInternalContextBase
*> finished_queue
;
237 // Dispatch, retry, queues
239 void inc_dispatch_depth() { ++dispatch_depth
; }
240 void dec_dispatch_depth() { --dispatch_depth
; }
241 void retry_dispatch(Message
*m
);
242 bool handle_deferrable_message(Message
*m
);
243 void _advance_queues();
244 bool _dispatch(Message
*m
, bool new_msg
);
246 ceph::heartbeat_handle_d
*hb
; // Heartbeat for threads using mds_lock
248 bool is_stale_message(Message
*m
) const;
250 map
<mds_rank_t
, version_t
> peer_mdsmap_epoch
;
252 ceph_tid_t last_tid
; // for mds-initiated requests (e.g. stray rename)
254 list
<MDSInternalContextBase
*> waiting_for_active
, waiting_for_replay
, waiting_for_reconnect
, waiting_for_resolve
;
255 list
<MDSInternalContextBase
*> replay_queue
;
256 map
<mds_rank_t
, list
<MDSInternalContextBase
*> > waiting_for_active_peer
;
257 map
<epoch_t
, list
<MDSInternalContextBase
*> > waiting_for_mdsmap
;
259 epoch_t osd_epoch_barrier
;
261 // Const reference to the beacon so that we can behave differently
266 * Emit clog warnings for any ops reported as warnings by optracker
268 void check_ops_in_flight();
270 int mds_slow_req_count
;
273 * Share MDSMap with clients
275 void bcast_mds_map(); // to mounted clients
276 epoch_t last_client_mdsmap_bcast
;
278 map
<mds_rank_t
,DecayCounter
> export_targets
; /* targets this MDS is exporting to or wants/tries to */
280 void create_logger();
283 void queue_waiter(MDSInternalContextBase
*c
) {
284 finished_queue
.push_back(c
);
285 progress_thread
.signal();
287 void queue_waiters(list
<MDSInternalContextBase
*>& ls
) {
288 finished_queue
.splice( finished_queue
.end(), ls
);
289 progress_thread
.signal();
295 LogChannelRef
&clog_
,
301 Context
*respawn_hook_
,
302 Context
*suicide_hook_
);
309 // Daemon lifetime functions: these guys break the abstraction
310 // and call up into the parent MDSDaemon instance. It's kind
311 // of unavoidable: if we want any depth into our calls
312 // to be able to e.g. tear down the whole process, we have to
313 // have a reference going all the way down.
320 * Call this periodically if inside a potentially long running piece
321 * of code while holding the mds_lock
323 void heartbeat_reset();
326 * Report state DAMAGED to the mon, and then pass on to respawn(). Call
327 * this when an unrecoverable error is encountered while attempting
328 * to load an MDS rank's data structures. This is *not* for use with
329 * errors affecting normal dirfrag/inode objects -- they should be handled
330 * through cleaner scrub/repair mechanisms.
332 * Callers must already hold mds_lock.
337 * Wrapper around `damaged` for users who are not
338 * already holding mds_lock.
340 * Callers must not already hold mds_lock.
342 void damaged_unlocked();
344 utime_t
get_laggy_until() const;
346 void send_message_mds(Message
*m
, mds_rank_t mds
);
347 void forward_message_mds(Message
*req
, mds_rank_t mds
);
349 void send_message_client_counted(Message
*m
, client_t client
);
350 void send_message_client_counted(Message
*m
, Session
*session
);
351 void send_message_client_counted(Message
*m
, Connection
*connection
);
352 void send_message_client_counted(Message
*m
, const ConnectionRef
& con
) {
353 send_message_client_counted(m
, con
.get());
355 void send_message_client(Message
*m
, Session
*session
);
356 void send_message(Message
*m
, Connection
*c
);
357 void send_message(Message
*m
, const ConnectionRef
& c
) {
358 send_message(m
, c
.get());
361 void wait_for_active_peer(mds_rank_t who
, MDSInternalContextBase
*c
) {
362 waiting_for_active_peer
[who
].push_back(c
);
364 void wait_for_cluster_recovered(MDSInternalContextBase
*c
) {
365 assert(cluster_degraded
);
366 waiting_for_active_peer
[MDS_RANK_NONE
].push_back(c
);
369 void wait_for_active(MDSInternalContextBase
*c
) {
370 waiting_for_active
.push_back(c
);
372 void wait_for_replay(MDSInternalContextBase
*c
) {
373 waiting_for_replay
.push_back(c
);
375 void wait_for_reconnect(MDSInternalContextBase
*c
) {
376 waiting_for_reconnect
.push_back(c
);
378 void wait_for_resolve(MDSInternalContextBase
*c
) {
379 waiting_for_resolve
.push_back(c
);
381 void wait_for_mdsmap(epoch_t e
, MDSInternalContextBase
*c
) {
382 waiting_for_mdsmap
[e
].push_back(c
);
384 void enqueue_replay(MDSInternalContextBase
*c
) {
385 replay_queue
.push_back(c
);
388 bool queue_one_replay();
390 void set_osd_epoch_barrier(epoch_t e
);
391 epoch_t
get_osd_epoch_barrier() const {return osd_epoch_barrier
;}
392 epoch_t
get_osd_epoch() const;
394 ceph_tid_t
issue_tid() { return ++last_tid
; }
398 MDSMap
*get_mds_map() { return mdsmap
; }
400 int get_req_rate() const { return logger
->get(l_mds_request
); }
402 int get_mds_slow_req_count() const { return mds_slow_req_count
; }
404 void dump_status(Formatter
*f
) const;
406 void hit_export_target(utime_t now
, mds_rank_t rank
, double amount
=-1.0);
407 bool is_export_target(mds_rank_t rank
) {
408 const set
<mds_rank_t
>& map_targets
= mdsmap
->get_mds_info(get_nodeid()).export_targets
;
409 return map_targets
.count(rank
);
412 bool evict_client(int64_t session_id
, bool wait
, bool blacklist
,
413 std::stringstream
& ss
, Context
*on_killed
=nullptr);
416 void dump_clientreplay_status(Formatter
*f
) const;
417 void command_scrub_path(Formatter
*f
, const string
& path
, vector
<string
>& scrubop_vec
);
418 void command_tag_path(Formatter
*f
, const string
& path
,
420 void command_flush_path(Formatter
*f
, const string
& path
);
421 void command_flush_journal(Formatter
*f
);
422 void command_get_subtrees(Formatter
*f
);
423 void command_export_dir(Formatter
*f
,
424 const std::string
&path
, mds_rank_t dest
);
425 bool command_dirfrag_split(
428 bool command_dirfrag_merge(
431 bool command_dirfrag_ls(
435 int _command_export_dir(const std::string
&path
, mds_rank_t dest
);
436 int _command_flush_journal(std::stringstream
*ss
);
437 CDir
*_command_dirfrag_get(
438 const cmdmap_t
&cmdmap
,
442 Messenger
*messenger
;
445 Context
*respawn_hook
;
446 Context
*suicide_hook
;
448 // Friended to access retry_dispatch
449 friend class C_MDS_RetryMessage
;
451 // FIXME the state machine logic should be separable from the dispatch
452 // logic that calls it.
454 void calc_recovery_set();
455 void request_state(MDSMap::DaemonState s
);
457 bool standby_replaying
; // true if current replay pass is in standby-replay mode
460 // The MDSMap is available, configure default layouts and structures
461 MDS_BOOT_INITIAL
= 0,
462 // We are ready to open some inodes
464 // We are ready to do a replay if needed
465 MDS_BOOT_PREPARE_LOG
,
466 // Replay is complete
469 friend class C_MDS_BootStart
;
470 friend class C_MDS_InternalBootStart
;
471 void boot_create(); // i am new mds.
472 void boot_start(BootStep step
=MDS_BOOT_INITIAL
, int r
=0); // starting|replay
475 void creating_done();
476 void starting_done();
478 void standby_replay_restart();
479 void _standby_replay_restart_finish(int r
, uint64_t old_read_pos
);
480 class C_MDS_StandbyReplayRestart
;
481 class C_MDS_StandbyReplayRestartFinish
;
485 void resolve_start();
487 void reconnect_start();
488 void reconnect_done();
489 void rejoin_joint_start();
492 void recovery_done(int oldstate
);
493 void clientreplay_start();
494 void clientreplay_done();
496 void stopping_start();
497 void stopping_done();
499 void validate_sessions();
503 void handle_mds_recovery(mds_rank_t who
);
504 void handle_mds_failure(mds_rank_t who
);
507 /* Update MDSMap export_targets for this rank. Called on ::tick(). */
508 void update_targets(utime_t now
);
511 /* This expects to be given a reference which it is responsible for.
512 * The finish function calls functions which
513 * will put the Message exactly once.*/
514 class C_MDS_RetryMessage
: public MDSInternalContext
{
518 C_MDS_RetryMessage(MDSRank
*mds
, Message
*m
)
519 : MDSInternalContext(mds
)
524 void finish(int r
) override
{
525 mds
->retry_dispatch(m
);
530 * The aspect of MDSRank exposed to MDSDaemon but not subsystems: i.e.
531 * the service/dispatcher stuff like init/shutdown that subsystems should
534 class MDSRankDispatcher
: public MDSRank
540 bool handle_asok_command(std::string command
, cmdmap_t
& cmdmap
,
541 Formatter
*f
, std::ostream
& ss
);
542 void handle_mds_map(MMDSMap
*m
, MDSMap
*oldmap
);
543 void handle_osd_map();
544 void update_log_config();
547 const cmdmap_t
&cmdmap
,
550 std::stringstream
*ds
,
551 std::stringstream
*ss
,
554 void dump_sessions(const SessionFilter
&filter
, Formatter
*f
) const;
555 void evict_clients(const SessionFilter
&filter
, MCommand
*m
);
557 // Call into me from MDS::ms_dispatch
558 bool ms_dispatch(Message
*m
);
563 LogChannelRef
&clog_
,
569 Context
*respawn_hook_
,
570 Context
*suicide_hook_
);
573 // This utility for MDS and MDSRank dispatchers.
574 #define ALLOW_MESSAGES_FROM(peers) \
576 if (m->get_connection() && (m->get_connection()->get_peer_type() & (peers)) == 0) { \
577 dout(0) << __FILE__ << "." << __LINE__ << ": filtered out request, peer=" << m->get_connection()->get_peer_type() \
578 << " allowing=" << #peers << " message=" << *m << dendl; \
584 #endif // MDS_RANK_H_